WorksComputer Science
EchoSpec: Training Search Agents to Weigh Evidence Above Echoes
Abstract
Search agents trained with reinforcement learning decide what to retrieve, when to search again, and when to answer. In real corpora, one claim is copied and paraphrased across many pages while the primary source settling a question appears once. We show that reinforcement-learned search agents grant echoes the credence of independent sources and lose their ability to abstain. Across five base models from 3B to 32B and seven released RL agents on multi-hop Wikipedia questions with primary evidence intact, eight paraphrased false-claim echoes overturn 58-73% of previously correct answers, while exact-match RL drives deleted-evidence abstention below 15%. Conditioning on retrieval localizes the failure to evidence acquisition: once primary evidence enters context, agents overrule eight echoes as reliably as one, whereas echoes fill top ranks and halt search early, so larger budgets, forced search turns, or passive passage filtering alone fail to recover displaced evidence. We introduce EchoSpec, specifying reliability across five corpus transformations with known target answers and training in two stages. Verification warm-start teaches multi-hop bridge-entity recovery past echoes, after which group RL optimizes an evidence-gated -reward with an outcome-independent acquisition bonus and scale-preserving advantages retaining the optimal rule . Across Qwen2.5-3B, Qwen2.5-7B, and Llama-3.1-8B backbones on nine evaluation suites, EchoSpec cuts eight-echo deference by 30-44 points on Qwen2.5-7B down to 29-31%, nearly quadruples eight-echo reach to 26-27% with zero-shot transfer to unseen BGE and BM25 retrievers, and raises standalone QA precision to 47-55%. Code is publicly available at https://github.com/unihsy/EchoSpec.
Figure 1. EchoSpec at a glance. Top: eight false echoes fill top results on a two-hop question, causing an exact-match RL agent to stop after one search and repeat the echo while EchoSpec looks up the bridge entity, reaches primary evidence, and abstains if it is deleted. Bottom: (a) Echoes flip correct answers across Qwen2.5 7B–32B and Search-R1. (b) Exact-match RL erases deleted-evidence abstention. (c) EchoSpec cuts deference at every echo count. (d) EchoSpec finds primary evidence under eight echoes while abstaining if evidence is deleted.
Introduction
Search agents trained with reinforcement learning decide for themselves what to look up, when to search again, and when to answer [4, 15, 31, 51]. Yet real search corpora are repetitive: syndicated news, aggregators, and machine-generated pages restate one claim many times while the primary source settling a question appears once. Repetition makes statements seem truer to humans [7, 12], and recent audits reveal the same vulnerability in language agents, which treat paraphrased copies of a false claim as independent corroboration [53] or abandon retrieved evidence under a distractor [25]. Meanwhile, reinforcement fine-tuning erodes the refusals that guard against unsupported answers in reasoning and search alike [32, 49].
We trace both failures directly to how search agents are trained. Across five base models from 3B to 32B and seven RL agents on multi-hop Wikipedia questions with primary evidence intact, eight paraphrased echoes of a false claim overturn 58–73% of answers correct on the clean corpus in Figure 2 and Table 3. Search-R1 defers more than its 7B base model at every echo count, and exact-match RL cuts abstention on deleted-evidence questions from 47% to below 15% in Figure 3. Both behaviors follow from standard RL training: clean QA corpora never separate repetition from evidence, and binary exact-match rewards pay agents for stopping at the first passage stating an answer while pricing guesses and refusals alike.
Figure 2. Echoes override intact evidence and end the search. Qwen2.5-Instruct at 7B, 14B, and 32B alongside Search-R1-7B on 1,114 multi-hop questions over wiki-18 across two held-out test sets, with primary evidence retrievable in every condition. (a) Share of clean-correct questions that switch to the injected answer as false echoes enter the corpus. (b) Accuracy on the clean corpus and under eight echoes of a true or false claim sharing identical text with the answer swapped. (c) Mean number of searches issued per question across the same three corpus conditions.
Figure 3. Exact-match rewards erase abstention under a prompt permitting abstention. (a) Qwen2.5-7B trained with the Search-R1 recipe sees deleted-evidence abstention collapse within 100 steps. (b) Search-R1-7B vs. Qwen2.5-7B across seven QA sets and deleted evidence. (c) Abstention with intact hollow markers vs. deleted filled markers.
| Method | Def@1 ↓ | Def@3 ↓ | Def@8 ↓ | Reach@8 ↑ | Acc@8 ↑ | Abst. ↑ | Prec. ↑ | RA-F1 ↑ |
| Released search agents | ||||||||
| Search-R1 | 46.0 | 63.3 | 69.1 | 13.3 | 6.8 | 0.2 | 38.4 | 38.4 |
| ZeroSearch | 53.0 | 69.6 | 72.9 | 7.3 | 3.8 | 0.1 | 35.2 | 35.2 |
| AutoRefine (Instruct) | 47.8 | 63.0 | 72.2 | 13.0 | 8.8 | 0.0 | 42.4 | 42.4 |
| AutoRefine (Base) | 40.5 | 57.2 | 59.8 | 18.1 | 12.5 | 0.0 | 44.7 | 44.7 |
| ReSearch | 43.5 | 64.5 | 68.9 | 20.9 | 9.4 | 0.0 | 40.6 | 40.6 |
| R1-Searcher | 33.3 | 54.2 | 61.9 | 21.0 | 12.5 | 0.0 | 41.8 | 41.7 |
| AWA-RL | 48.4 | 55.6 | 58.6 | 19.8 | 6.0 | 0.0 | 35.9 | 35.9 |
| Qwen2.5-7B with inference-time defenses | ||||||||
| No defense | 41.9 | 57.2 | 58.8 | 7.1 | 5.5 | 49.2 | 41.1 | 33.8 |
| Defensive prompt | 42.1 | 58.3 | 61.0 | 7.9 | 4.7 | 52.7 | 42.0 | 32.9 |
| Near-duplicate removal | 41.3 | 55.9 | 63.0 | 7.1 | 4.8 | 46.7 | 40.6 | 33.6 |
| Provenance gate | 44.9 | 56.1 | 59.4 | 7.6 | 5.8 | 46.8 | 41.0 | 33.9 |
| TrustRAG | 35.4 | 41.7 | 22.8 | 11.4 | 9.4 | 48.0 | 40.3 | 33.0 |
| MADAM-RAG† | 7.4 | 40.6 | 48.4 | 7.0 | 2.8 | 84.4 | 55.7 | 28.3 |
| Qwen2.5-7B trained on the same questions | ||||||||
| Exact-match RL | 35.2 | 59.5 | 67.7 | 15.9 | 8.1 | 1.4 | 37.2 | 36.8 |
| -reward RL | 45.4 | 68.4 | 69.4 | 9.4 | 5.2 | 32.5 | 42.4 | 37.6 |
| EchoSpec (3 seeds) | 14.3 | 23.9 | 28.9 | 25.8 | 16.7 | 37.6 | 47.1 | 40.1 |
| Llama-3.1-8B cross-family replication (1 seed) | ||||||||
| No defense | 37.3 | 67.5 | 70.0 | 10.8 | 3.8 | 5.0 | 28.4 | 27.6 |
| Exact-match RL | 33.8 | 61.3 | 66.7 | 17.6 | 9.8 | 2.8 | 42.7 | 42.2 |
| EchoSpec | 11.1 | 24.1 | 30.6 | 26.5 | 17.2 | 52.5 | 55.4 | 44.2 |
| Qwen2.5-7B with orthogonal retrieval filters (3 seeds) | ||||||||
| EchoSpec + TrustRAG | 13.8 | 19.4 | 12.8 | 29.4 | 19.0 | 36.5 | 44.1 | 38.3 |
| EchoSpec + Sem. dedup | 16.7 | 25.9 | 28.7 | 36.2 | 20.2 | 33.9 | 46.2 | 40.0 |
Table 3. Main results across two ten-world suites (1,114 items) and seven QA sets (3-seed Qwen2.5-7B means; 1-seed Llama-3.1-8B). Def@m: conditional deference at echoes (↓). Reach@8, Acc@8: evidence reach and EM at (↑). Abst.: Delete abstention (↑). Prec., RA-F1: QA precision and RA-F1 [49] (↑). † MADAM-RAG abstains on 76% of clean items (2.8% Acc@8). Standalone 1st/2nd and filter rows beating standalone are shaded.
Conditioning on retrieval localizes echo deference to evidence acquisition. Once the primary passage enters the context, agents overrule eight echoes as reliably as one, deferring on 25–34% of clean-correct items at eight echoes versus 27–39% at one. Echoes succeed by keeping primary evidence out of view: they fill top retrieval ranks, halt search after one or two queries, and slash primary-evidence reach from 53–72% to 20–31%. Defenses that filter or debate retrieved passages [36, 52, 53] act only on documents already fetched, whereas recovering displaced evidence requires changing the search policy itself.
We introduce EchoSpec, a training framework that builds active verification and calibrated abstention into the search policy. EchoSpec defines reliability across five corpus transformations with known target answers: duplicating primary evidence or injecting echoes of a true or false claim preserves the answer, rewriting primary evidence updates it, and deleting primary evidence requires ABSTAIN. Because true and false echoes share the exact generated text with only the answer swapped, neither style nor copy count reveals veracity. Stage 1 warms up the policy on grounded multi-hop verification trajectories so the agent learns to query bridge entities past Turn-1 echoes. Stage 2 runs group RL with an evidence-gated reward scoring +1 when grounded or rightly abstaining, 0 on answerable refusals, and when wrong, with an acquisition bonus preserving and scale-preserving advantages preventing collapse.
We evaluate EchoSpec across nine evaluation suites: two held-out multi-hop suites built with Qwen3-32B [46] and Phi-4 [1], each spanning ten corpus worlds per question across Identity, Dup(3), Inject(1, 3, 8), Consistent(1, 3, 8), Rewrite, and Delete, plus seven standard QA benchmarks. Across Qwen2.5-3B, Qwen2.5-7B [47], and Llama-3.1-8B [11], EchoSpec cuts eight-echo deference to 29–31%, beating all seven released RL agents on Qwen2.5-7B by 30–44 points while transferring zero-shot to unseen BGE [42] and BM25 [28] retrievers at 25–30% deference. It simultaneously lifts eight-echo evidence reach nearly fourfold to 26–27% and raises QA precision to 47–55%, yielding four main contributions:
A controlled audit localizing echo deference in RL search agents to premature search termination and showing that outcome-reward RL amplifies deference while eroding abstention on unsupported inputs.
Five corpus-transformation specifications with symmetric claim construction and complete evidence deletion, instantiated as training worlds and two evaluation suites generated by distinct model families.
EchoSpec, combining a multi-hop verification warm-start with -calibrated group RL, an outcome-independent evidence-acquisition bonus, and scale-preserving advantage estimation.
Evaluation across Qwen2.5 and Llama-3.1 backbones, seven RL agents, five defenses, forced multi-turn baselines, and unseen dense and BM25 retrievers across nine suites, showing consistent gains.
| Work | Intervention | Agentic | Echoes | Recovers | Trained | Abstains | Template-free |
| RobustRAG [41] | isolate and aggregate | ||||||
| TrustRAG [52] | cluster filtering | ||||||
| MADAM-RAG [36] | multi-agent debate | ||||||
| Evidence Laundering [53] | provenance gate | ||||||
| DRNOISE [25] | audit only | ||||||
| Search-R1 family [15, 30, 31] | outcome-reward RL | ||||||
| QREF [5] | query refinement | ||||||
| Hallucination Tax [32] | unanswerable data | ||||||
| AWA-RL [49] | abstention reward | ||||||
| EchoSpec (ours) | warm-start, -RL |
Table 1. Positioning against prior work. Agentic: studies a multi-turn search policy. Echoes: handles many copies of one false claim. Recovers: brings displaced evidence back into context, beyond filtering what was retrieved. Trained: changes the policy through training. Abstains: models abstention when evidence is missing. Template-free: needs no shared wording, embedding similarity, or provenance metadata among copies. marks partial coverage.
Related Work
Reinforcement-learned search agents. Building on prompted retrieval pipelines [3, 35], Search-R1 [15], R1-Searcher [31], ReSearch [4], DeepResearcher [51], ZeroSearch [33], and AutoRefine [30] train search policies end to end with outcome rewards, typically exact match on multi-hop QA [34, 48] over a fixed Wikipedia index [20, 37]. Later work refines queries and filters evidence to cut redundant searches [5]. EchoSpec keeps the standard training protocol and changes what the agent learns from: corpus variants with known target answers, a verification warm-start, and a reward pricing wrong guesses.
Echoes, poisoning, and conflicting evidence. Retrieval corpora can be poisoned with a handful of crafted passages [26, 54], and models weigh such context against their parametric memory [23, 40, 43, 44]. Closest to our setting, Zhou et al. [53] show that frozen deep research agents treat paraphrased copies of one false claim as independent support, and that a provenance gate collapsing pages with a shared writing skeleton reduces the effect. DRNOISE [25] shows that one misleading document can override evidence that requires multi-step derivation, and graph agents revisit redundant evidence paths [38]. Defenses act on the retrieved set: provenance gating [53], clustering away injected passages [52], isolate-then-aggregate decoding [41], multi-agent debate over retrieved documents [36], and retrieval correction [45]. EchoSpec trains the search policy that decides what enters the retrieved set, on echoes with no shared template, and studies how reinforcement learning shapes this policy in open search agents trained at scale with outcome rewards.
Abstention and selective prediction. Selective prediction trades coverage for risk through a confidence threshold [6, 9, 19]. Language models carry usable signals of their own correctness [17], which training can sharpen into refusals [8, 50], as surveyed by Wen et al. [39]. Kalai et al. [18] argue that binary grading rewards guessing over expressing uncertainty, reinforcement fine-tuning measurably lowers refusal on unanswerable questions [32], and AWA-RL shapes an abstention reward for search agents from the model’s prior capability [49]. EchoSpec prices wrong answers with a single parameter that has a closed-form answering threshold, identifies standard advantage normalization as the step that erases this threshold during training, and couples abstention with a search policy that actively seeks out the primary evidence first.
Learning under controlled transformations. Counterfactually augmented data pairs inputs that differ only in the feature that should change the label [21], and invariance principles train predictors whose behavior is stable across environments [2], both aimed at the shortcut features of standard training data [10]. EchoSpec brings this idea to agents acting over a corpus: each transformation has a specified effect on the answer, and true and false claims share their wording, so no stylistic shortcut reveals the injected text.
Reinforcement-Learned Search Agents Follow Echoes
Frozen research agents treat paraphrased copies of a false claim as independent sources [53], and reinforcement fine-tuning weakens refusal [32, 49]. We measure both effects on reinforcement-learned search agents over the corpus they train on and locate where in the search process the echoes take hold.
Setup. We follow the Search-R1 protocol [15]: the 21M-passage wiki-18 corpus [20], the e5-base-v2 retriever [37], top-3 passages per query, and at most four searches, reproducing released Search-R1 at 38.7 average exact match against 38.5 reported. Test questions come from the HotpotQA [48] and MuSiQue [34] development sets, each with an identified passage that states the answer. A generator writes a short claim stating the answer without support, and the false claim swaps in a plausible alternative answer across title and body. An Inject() world adds paraphrased echoes of the false claim and a Consistent() world adds echoes of the true claim. Generators from two model families, Qwen3-32B [46] and Phi-4 [1], build test sets of 507 and 607 questions on which a TF-IDF classifier separates true from false echoes at AUC 0.60–0.68. We evaluate Qwen2.5-Instruct [47] at 7B, 14B, and 32B with a prompt permitting ABSTAIN, and Search-R1-7B with its official prompt. Conditional deference is the share of clean-correct questions switching to the injected answer under echoes, with full world-construction details in Appendix § 7.
Echoes override evidence
Deference grows with the number of echoes. A single echo overturns 29–48% of correct answers, and eight echoes overturn 58–66% across Qwen2.5 7B–32B and Search-R1 in Figure 2a, reaching 62.5% on Qwen2.5-3B and 70.0% on Llama-3.1-8B [11]. Deference rises with for every agent on both suites, and model scale leaves it intact: 32B defers on 63% of clean-correct answers at versus 61% at 7B.
Reinforcement learning amplifies deference. Search-R1 defers the most at every , reaching 48% at and 66% at , four to five points above the 7B model it was trained from. Training the same base model with the Search-R1 recipe on our multi-hop training questions reproduces the effect, with deference at reaching 63–69% across two seeds. Outcome rewards on clean corpora pay the agent for answering from the first passage that states an answer, and echoes supply such passages in quantity at the top of the ranking.
Echoes set the answer and end the search. Eight echoes of the true claim raise accuracy from 17–26% to 48–57%, while eight echoes of the false claim lower it to 5–9% in Figure 2b, with the same evidence retrievable throughout. Both kinds of echo cut search effort by the same amount in Figure 2c. Search-R1 drops from 2.49 searches per question to 1.89 and 1.91 because the agent stops once a passage states an answer.
The failure lies in evidence acquisition
Search agents weigh evidence well once they have it. When a trajectory retrieves the primary evidence, deference at eight echoes is 25–34% in Table 2, no higher than the 27–39% at a single echo. When the evidence stays out of the context, deference reaches 70–85% at eight echoes. The echoes act on what the agent sees: they fill the top retrieval results and end the search early, and the share of trajectories that reach the primary evidence falls from 53–72% at one echo to 20–31% at eight. All three agents show this pattern. Robustness to echoes therefore requires a search policy that continues past the echoes until it reaches the primary evidence, a behavior that clean training corpora never reward in any question.
| Agent | Evidence reach (%) | Evidence reach (%) | Evidence reach (%) | Deference, retrieved (%) | Deference, retrieved (%) | Deference, retrieved (%) | Deference, missed (%) | Deference, missed (%) | Deference, missed (%) |
| Qwen2.5-7B | 52.6 | 26.3 | 20.1 | 39.2 | 31.4 | 30.8 | 52.2 | 67.1 | 69.7 |
| Exact-match RL | 72.0 | 42.7 | 31.2 | 27.0 | 42.5 | 33.7 | 56.8 | 72.2 | 81.9 |
| -reward RL | 65.4 | 31.7 | 22.1 | 31.2 | 27.6 | 24.5 | 66.3 | 80.5 | 84.5 |
Table 2. Echoes act by keeping the evidence out of reach. Questions each agent answers correctly on the clean corpus, split by whether it retrieved the primary evidence. Reach is the share of trajectories that retrieved it.
Exact-match rewards remove abstention
Exact match scores a wrong answer and an abstention identically on answerable-only training sets. Under the reward of Section 4.3, this corresponds to , whose optimal answering threshold in Proposition 1 is zero: any guess with nonzero match probability beats abstaining, explaining refusal loss under RL fine-tuning [32, 49]. Training Qwen2.5-7B with Search-R1 drops abstention on held-out deleted-evidence questions from 47% to 3–14% across three seeds within 100 steps in Figure 3a. Likewise, released Search-R1-7B abstains 5–15 points less than its base model across all seven QA sets in Figure 3b, and every tested model abstains only 5–10 points more often when primary evidence is deleted than when intact in Figure 3c.
EchoSpec
EchoSpec trains a search agent to pursue primary evidence past echoes and abstain when support is absent. It specifies target behavior across five corpus transformations in § 4.1, seeds multi-step verification through a trajectory warm-start in § 4.2, and optimizes calibrated answering and search depth with threshold-preserving group RL in Section 4.3, as summarized in Figure 4 for a multi-hop training question.
Figure 4. Overview of EchoSpec. (a) Controlled worlds & search: seed defines answer-invariant Identity, Dup, Inject, and Consistent worlds linked by , plus answer-shifting Rewrite and Delete under -query top-3 search masking seen passages. (b) Stage 1 · Verification warm-start: temporary chain hints guide a 32B solver past Turn-1 echoes via to , and three filters keep grounded, self-contained rollouts for LoRA SFT . (c) Stage 2 · -calibrated group RL: from with KL reference , optimizes on paired and rollouts via with without , keeping .
Reliability as a specification over corpus transformations
Each training question comes with gold answer , bridge articles , and primary passage stating . A transformation fixes a known target output . Answer-invariant worlds set : Dup adds paraphrases of , Inject adds echoes of a bare claim stating decoy , and Consistent adds echoes of the identical claim template stating . Answer-shifting worlds modify support: Rewrite minimally edits to state , and Delete masks with target , verified iteratively by probing the retriever with , , and until no unmasked passage surfaces . Because true and false echoes differ solely in swapping across title and body, neither wording, length, nor copy count reveals veracity, so satisfying all worlds requires checking claims against primary evidence.
Verification warm-start
Policy-gradient RL can reinforce only behaviors with nonzero sampling mass, yet § 3.2 shows that base agents rarely search past Turn-1 echoes. Stage 1 seeds this verification behavior via chain-guided trajectory synthesis over worlds weighted toward multi-echo Inject and Delete. During synthesis only, a note appended to the prompt names the supporting article chain and instructs the solver to query entities from rather than bare claim answers until confirming in or exhausting search. Default synthesis uses Qwen2.5-32B-Instruct, which itself defers at 63% under eight echoes without guidance, while self-guided 7B synthesis is ablated in § 5.2. Retained trajectories satisfy three quality criteria: (i) grounded outcome, requiring with or -search abstention on DELETE, (ii) hop continuity, requiring every hinted title , absent from to appear in an earlier retrieved passage before being queried, and (iii) self-contained reasoning, rejecting any mention of the guidance note. Prioritizing -search echo recoveries and stripping the note yields plain-prompt demonstrations for LoRA [14] fine-tuning.
-calibrated group reinforcement learning
Decision-theoretic reward. In world , terminal output over retrieved passages receives reward
where discounts unverified echo matches on CONSISTENT worlds relative to grounded answers, and standard exact-match RL corresponds to the degenerate special case and of Eq. (1).
Proposition 1. Fix retrieved passages on an answerable world with , and let be the conditional probability of given . Under Eq. (1), answering beats abstaining in expectation iff .
Proof. Expected return is for answering versus 0 for ABSTAIN. □
Scale-preserving advantages. Proposition 1 characterizes the Bayes-optimal stopping payoff for fixed rather than guaranteeing calibrated posteriors after finite-sample RL. However, standard GRPO [29] erases from the objective by dividing centered rewards by their group standard deviation:
Proposition 2. In any failed group of rollouts with abstentions at and wrong answers at for , dividing centered returns by the group standard deviation assigns advantage to abstentions and to wrong answers identically for every risk penalty .
Proof. Group mean and standard deviation scale in and cancel. □
Empirically, std-normalized GRPO at drives answerable-world abstention to 61–79% within 50 steps, whereas mean-centered advantages keep rewards in units of for monotonic control in Appendix § 10.
Outcome-independent acquisition bonus and cross-world baseline. To reward multi-hop exploration without shifting the abstention rule, we add a terminal set bonus to each episode return, where . Unlike step-wise potential shaping, depends only on retrieved set , not , so it cancels from for every and leaves Proposition 1’s payoff rule intact. Each step also rolls out episodes in and in . On invariant worlds , , and , optional cross-world baseline serves as an unbiased control variate breaking uniform ties in Proposition 4, with once Stage 1 prevents collapse.
Search protocol and inference-time modularity
At training and test time, the agent queries e5-base-v2 over wiki-18 for top-3 passages in up to four searches without source metadata. Within each episode, exclude-seen retrieval masks prior passage IDs so reformulated queries surface fresh documents rather than repeating top echoes, lifting eight-echo reach while leaving deference resistance intact when disabled in § 5.2 and Appendix § 12. Because EchoSpec updates only policy weights, it deploys directly and composes with retrieval filters in § 5.1.
Evaluation
Setup and main results
Setup. Suites span Test A/B (Appendix § 10) and seven QA sets (Appendix § 11): NQ [22], TriviaQA [16], PopQA [24], HotpotQA [48], 2WikiMultiHopQA [13], MuSiQue [34], and Bamboogle [27]. RL agents: Search-R1 [15], ZeroSearch [33], AutoRefine-Base/Instruct [30], ReSearch [4], R1-Searcher [31], and AWA-RL [49]. On Qwen2.5-7B [47], defenses cover prompts, dedup, skeleton gating [53], TrustRAG [52], and MADAM-RAG [36]; we train Exact-match RL, -reward RL, and EchoSpec on Qwen2.5-3B/7B and Llama-3.1-8B [11] with protocol, BGE [42], and BM25 [28] controls (Appendices § 12 and § 13).
Active evidence recovery, cross-backbone, and cross-retriever transfer. On Qwen2.5-7B, standalone EchoSpec cuts deference to , , and in Table 3, beating all seven RL agents (58.6–72.9%) by 30–44 points and lifting Reach@8 from 7.1% to and Acc@8 to . Across backbones (Table 9), EchoSpec cuts Def@8 from 62.5% to 30.4% on Qwen2.5-3B and from 70.0% to 30.6% on Llama-3.1-8B at 26.5% Reach@8. Controls use protocol artifacts: exclude-seen (Table 8) leaves RL agents at 59.7–74.7% Def@8, forcing three Search-R1 turns (Table 10) gives 3.18 searches at 55.5% Def@8, and AWA-RL at runs 3.65 at 61.8%. Frozen EchoSpec transfers zero-shot (Table 11) to BGE at 29.4% and BM25 at 25.1% Def@8, and filters reach 11.8–12.8% Def@8 or 36.2–36.4% Reach@8.
| Method | Def@1 | Def@3 | Def@8 | Reach@8 | Acc@8 | Abst. | Prec. | RA-F1 |
| Released RL search agents under unified exclude-seen retrieval | ||||||||
| Search-R1 (+ exclude-seen) | 40.3 | 59.6 | 64.4 | 14.0 | 7.3 | 0.3 | 37.3 | 37.3 |
| Search-R1, abstention prompt (+ exclude-seen) | 42.8 | 59.8 | 66.0 | 11.8 | 6.1 | 34.7 | 42.7 | 37.8 |
| ZeroSearch (+ exclude-seen) | 55.4 | 70.4 | 74.7 | 7.4 | 4.3 | 0.2 | 35.1 | 35.1 |
| AutoRefine, Instruct (+ exclude-seen) | 44.4 | 63.3 | 71.8 | 13.5 | 8.3 | 0.1 | 42.9 | 42.9 |
| AutoRefine, Base (+ exclude-seen) | 36.3 | 56.0 | 59.7 | 18.2 | 12.9 | 0.0 | 45.3 | 45.3 |
| ReSearch (+ exclude-seen) | 32.0 | 56.4 | 70.6 | 21.6 | 8.9 | 0.0 | 40.4 | 40.4 |
| R1-Searcher (+ exclude-seen) | 24.1 | 46.1 | 60.7 | 22.3 | 13.3 | 0.0 | 40.8 | 40.7 |
| AWA-RL (+ exclude-seen) | 35.9 | 45.0 | 60.3 | 20.7 | 5.9 | 0.0 | 34.3 | 34.3 |
| Qwen2.5-7B defenses under unified exclude-seen retrieval | ||||||||
| Qwen2.5-7B, no defense (+ exclude-seen) | 41.0 | 57.8 | 60.1 | 7.6 | 5.7 | 48.5 | 41.4 | 34.3 |
| Defensive prompt (+ exclude-seen) | 38.9 | 54.5 | 62.7 | 8.1 | 4.4 | 52.2 | 41.1 | 32.4 |
| Near-duplicate removal (+ exclude-seen) | 41.9 | 57.3 | 61.4 | 8.0 | 5.9 | 47.3 | 41.0 | 33.9 |
| Provenance gate (+ exclude-seen) | 43.3 | 54.0 | 64.1 | 7.0 | 5.2 | 47.8 | 40.8 | 33.7 |
| TrustRAG (+ exclude-seen) | 40.2 | 38.9 | 24.6 | 10.8 | 9.5 | 49.8 | 40.8 | 33.6 |
| EchoSpec with and without exclude-seen retrieval | ||||||||
| Qwen2.5-7B + EchoSpec (w/o exclude-seen, 3 seeds) | 17.0 | 28.7 | 31.6 | 22.5 | 15.2 | 40.3 | 47.9 | 40.1 |
| Qwen2.5-7B + EchoSpec (w/o exclude-seen train & eval) | 15.9 | 26.1 | 27.1 | 23.6 | 14.3 | 50.1 | 50.2 | 40.3 |
| Qwen2.5-7B + EchoSpec (+ exclude-seen, 3 seeds) | 14.3 | 23.9 | 28.9 | 25.8 | 16.7 | 37.6 | 47.1 | 40.1 |
| Llama-3.1-8B + EchoSpec (+ exclude-seen, 1 seed) | 11.1 | 24.1 | 30.6 | 26.5 | 17.2 | 52.5 | 55.4 | 44.2 |
Table 8. Unified retrieval-protocol fairness control comparing all seven released RL search agents and five Qwen2.5-7B inference defenses with episode-level exclude-seen enabled against EchoSpec evaluated both with and without exclude-seen across all nine evaluation suites, highlighting standalone 1st and 2nd in each column.
| Backbone and Method | Def@1 ↓ | Def@3 ↓ | Def@8 ↓ | Reach@8 ↑ | Acc@8 ↑ | Abst. ↑ | Prec. ↑ | RA-F1 ↑ |
| Qwen2.5-3B-Instruct (3B parameters, cross-scale, 1 seed) | ||||||||
| No defense | 40.7 | 67.0 | 62.5 | 5.1 | 2.8 | 36.3 | 26.0 | 22.9 |
| No defense (+ exclude-seen) | 44.5 | 61.5 | 65.2 | 4.7 | 2.4 | 37.0 | 26.4 | 23.3 |
| TrustRAG | 38.9 | 39.9 | 27.2 | 7.8 | 4.4 | 37.6 | 25.7 | 22.1 |
| Exact-match RL | 42.7 | 69.4 | 71.5 | 11.6 | 5.3 | 5.2 | 35.1 | 34.6 |
| Warm-start only | 18.3 | 35.5 | 39.1 | 19.3 | 11.7 | 25.2 | 40.7 | 36.7 |
| EchoSpec | 15.3 | 29.6 | 30.4 | 22.5 | 12.4 | 37.5 | 42.8 | 36.9 |
| EchoSpec + TrustRAG | 15.8 | 23.5 | 12.3 | 23.5 | 15.6 | 38.6 | 42.1 | 36.4 |
| EchoSpec + Sem. dedup | 20.6 | 34.6 | 34.5 | 30.1 | 14.5 | 36.6 | 42.7 | 36.9 |
| Qwen2.5-7B-Instruct (7B parameters, 3-seed EchoSpec) | ||||||||
| No defense | 41.9 | 57.2 | 58.8 | 7.1 | 5.5 | 49.2 | 41.1 | 33.8 |
| No defense (+ exclude-seen) | 41.0 | 57.8 | 60.1 | 7.6 | 5.7 | 48.5 | 41.4 | 34.3 |
| TrustRAG | 35.4 | 41.7 | 22.8 | 11.4 | 9.4 | 48.0 | 40.3 | 33.0 |
| Exact-match RL (1 seed) | 35.2 | 59.5 | 67.7 | 15.9 | 8.1 | 1.4 | 37.2 | 36.8 |
| Warm-start only (3 seeds) | 20.8 | 35.0 | 47.2 | 21.0 | 13.9 | 33.1 | 47.0 | 41.5 |
| EchoSpec (3 seeds) | 14.3 | 23.9 | 28.9 | 25.8 | 16.7 | 37.6 | 47.1 | 40.1 |
| EchoSpec + TrustRAG (3 seeds) | 13.8 | 19.4 | 12.8 | 29.4 | 19.0 | 36.5 | 44.1 | 38.3 |
| EchoSpec + Sem. dedup (3 seeds) | 16.7 | 25.9 | 28.7 | 36.2 | 20.2 | 33.9 | 46.2 | 40.0 |
| Llama-3.1-8B-Instruct (8B parameters, cross-family, 1 seed) | ||||||||
| No defense | 37.3 | 67.5 | 70.0 | 10.8 | 3.8 | 5.0 | 28.4 | 27.6 |
| No defense (+ exclude-seen) | 34.7 | 71.8 | 74.4 | 11.1 | 4.2 | 4.2 | 26.6 | 25.9 |
| TrustRAG | 37.6 | 52.1 | 28.2 | 17.1 | 7.6 | 4.8 | 27.7 | 26.8 |
| Exact-match RL | 33.8 | 61.3 | 66.7 | 17.6 | 9.8 | 2.8 | 42.7 | 42.2 |
| Warm-start only | 22.7 | 35.2 | 45.7 | 21.0 | 14.6 | 24.7 | 45.5 | 42.3 |
| EchoSpec | 11.1 | 24.1 | 30.6 | 26.5 | 17.2 | 52.5 | 55.4 | 44.2 |
| EchoSpec + TrustRAG | 11.2 | 19.2 | 11.8 | 28.7 | 20.8 | 52.8 | 55.3 | 43.6 |
| EchoSpec + Sem. dedup | 17.1 | 26.9 | 26.6 | 36.4 | 20.9 | 54.2 | 55.9 | 44.4 |
Table 9. Cross-scale and cross-family replication across Qwen2.5-3B-Instruct (1 seed), Qwen2.5-7B-Instruct (3 seeds), and Llama-3.1-8B-Instruct (1 seed) on the two generator-disjoint suites and seven QA benchmarks, highlighting standalone 1st and 2nd alongside filter compositions that surpass the best standalone agent.
| Agent and Budget | Srch@0 | Srch@8 | Id. Acc ↑ | Def@1 ↓ | Def@3 ↓ | Def@8 ↓ | Reach@8 ↑ | Acc@8 ↑ | Abst. ↑ |
| Search-R1 | |||||||||
| , (default) | 2.52 | 1.94 | 26.3 | 46.0 | 63.3 | 69.1 | 13.3 | 6.8 | 0.2 |
| , | 2.65 | 2.00 | 27.2 | 48.9 | 66.4 | 65.7 | 13.7 | 7.2 | 0.3 |
| , | 2.71 | 2.02 | 26.5 | 50.1 | 62.0 | 66.2 | 13.8 | 7.9 | 0.1 |
| , (+ exclude-seen) | 2.82 | 2.03 | 26.6 | 44.0 | 58.5 | 67.4 | 14.1 | 7.0 | 0.3 |
| , | 2.36 | 1.80 | 26.5 | 49.9 | 62.6 | 72.1 | 14.5 | 6.7 | 0.1 |
| AutoRefine (Base) | |||||||||
| , (default) | 2.22 | 1.90 | 33.0 | 40.5 | 57.2 | 59.8 | 18.1 | 12.5 | 0.0 |
| , | 2.24 | 1.90 | 32.7 | 39.1 | 57.7 | 61.3 | 17.9 | 12.2 | 0.0 |
| , | 2.24 | 1.90 | 33.3 | 39.7 | 57.1 | 61.1 | 18.0 | 12.8 | 0.0 |
| , (+ exclude-seen) | 2.23 | 1.91 | 33.0 | 38.0 | 56.5 | 59.8 | 18.8 | 13.1 | 0.0 |
| , | 2.16 | 1.86 | 33.9 | 33.6 | 51.1 | 63.4 | 21.2 | 11.2 | 0.0 |
| AWA-RL | |||||||||
| , (default) | 3.19 | 3.01 | 24.3 | 48.4 | 55.6 | 58.6 | 19.8 | 6.0 | 0.0 |
| , | 3.86 | 3.38 | 28.7 | 43.9 | 57.4 | 63.0 | 19.6 | 7.5 | 0.2 |
| , | 4.26 | 3.55 | 30.4 | 49.5 | 60.0 | 67.2 | 21.9 | 6.4 | 0.5 |
| , (+ exclude-seen) | 4.38 | 3.65 | 29.2 | 38.4 | 53.3 | 61.8 | 25.6 | 8.1 | 0.6 |
| , | 3.13 | 2.86 | 24.8 | 42.4 | 54.7 | 63.6 | 23.1 | 5.7 | 0.0 |
| Forced multi-turn verification baselines (, ) | |||||||||
| Qwen2.5-7B + verify prompt (+ xs) | 1.80 | 1.46 | 20.3 | 40.7 | 62.3 | 64.8 | 10.6 | 5.1 | 34.7 |
| Qwen2.5-7B + verify, min 3 turns (+ xs) | 3.01 | 3.01 | 21.1 | 26.7 | 36.5 | 54.1 | 27.5 | 7.8 | 32.1 |
| Search-R1, min 3 turns (+ exclude-seen) | 3.35 | 3.18 | 24.7 | 20.2 | 36.0 | 55.5 | 29.5 | 11.0 | 0.1 |
| Search-R1 + verify, min 3 turns (+ xs) | 3.30 | 3.19 | 23.4 | 18.9 | 35.7 | 58.7 | 28.1 | 9.8 | 22.4 |
| AutoRefine (Base), min 3 turns (+ xs) | 2.28 | 1.95 | 33.1 | 32.1 | 49.9 | 53.8 | 19.7 | 14.3 | 0.0 |
| EchoSpec (Qwen2.5-7B, 3 seeds) | |||||||||
| , (+ exclude-seen, default) | 2.71 | 2.81 | 27.3 | 14.3 | 23.9 | 28.9 | 25.8 | 16.7 | 37.6 |
| , (+ exclude-seen) | 2.92 | 3.02 | 28.6 | 12.6 | 23.1 | 31.1 | 28.0 | 17.7 | 43.2 |
| , (+ exclude-seen) | 3.00 | 3.07 | 27.5 | 13.5 | 22.1 | 29.8 | 28.1 | 18.3 | 43.6 |
| , (+ exclude-seen) | 2.60 | 2.65 | 28.7 | 13.9 | 17.1 | 29.1 | 34.8 | 20.2 | 34.9 |
Table 10. Search horizon , passage depth , and forced multi-turn verification across Test Sets A and B with 3-seed EchoSpec means. Srch@0 and Srch@8 give mean searches on Identity and eight-echo Inject worlds, and Id. Acc gives clean accuracy, highlighting 1st and 2nd in each metric column.
| Retriever and Method | Srch@0 | Srch@8 | Id. Acc | Def@1 | Def@3 | Def@8 | Reach@8 | Acc@8 | Abst. |
| Zero-shot transfer to unseen dense retriever: BAAI/bge-base-en-v1.5 | |||||||||
| Search-R1 | 2.56 | 1.97 | 26.7 | 53.3 | 65.3 | 72.2 | 11.6 | 6.1 | 0.3 |
| Search-R1 (+ exclude-seen) | 2.60 | 1.97 | 27.0 | 45.9 | 62.6 | 71.0 | 12.3 | 6.3 | 0.3 |
| AutoRefine (Base) | 2.20 | 1.87 | 31.7 | 41.3 | 62.4 | 65.3 | 13.2 | 10.3 | 0.0 |
| AutoRefine, Base (+ exclude-seen) | 2.19 | 1.88 | 32.1 | 37.3 | 59.8 | 66.1 | 13.4 | 10.2 | 0.0 |
| R1-Searcher | 2.51 | 2.27 | 31.5 | 34.5 | 57.0 | 61.3 | 20.9 | 11.6 | 0.1 |
| R1-Searcher (+ exclude-seen) | 2.52 | 2.29 | 30.8 | 22.7 | 50.5 | 62.2 | 20.8 | 11.5 | 0.0 |
| AWA-RL (+ exclude-seen) | 3.19 | 3.02 | 23.9 | 27.8 | 43.5 | 58.1 | 21.2 | 5.6 | 0.2 |
| Qwen2.5-7B (no defense) | 1.22 | 1.10 | 14.8 | 40.8 | 58.9 | 61.7 | 6.1 | 4.7 | 48.5 |
| Qwen2.5-7B + TrustRAG | 1.23 | 1.21 | 13.5 | 39.3 | 49.8 | 20.7 | 9.8 | 8.2 | 49.0 |
| Qwen2.5-7B + Exact-match RL (1 seed) | 2.51 | 2.01 | 25.8 | 37.6 | 58.4 | 67.7 | 14.3 | 7.6 | 1.3 |
| Qwen2.5-7B + EchoSpec (+ xs, 3 seeds) | 2.80 | 2.92 | 26.9 | 13.4 | 21.2 | 29.4 | 26.0 | 15.2 | 36.3 |
| Qwen2.5-7B + EchoSpec (w/o xs, 1 seed) | 2.71 | 2.84 | 27.6 | 19.6 | 31.8 | 34.3 | 21.5 | 14.2 | 33.7 |
| Llama-3.1-8B (no defense) | 3.04 | 2.16 | 17.2 | 31.1 | 67.4 | 67.2 | 11.2 | 4.1 | 6.0 |
| Llama-3.1-8B + EchoSpec (+ xs, 1 seed) | 2.60 | 2.81 | 29.3 | 11.3 | 20.8 | 30.5 | 25.2 | 15.4 | 51.8 |
| Zero-shot transfer to lexical sparse retriever: Okapi BM25 | |||||||||
| Search-R1 | 2.75 | 1.94 | 22.8 | 49.2 | 63.1 | 64.1 | 9.9 | 6.1 | 0.3 |
| Search-R1 (+ exclude-seen) | 2.81 | 1.94 | 22.4 | 38.2 | 58.6 | 64.8 | 11.9 | 6.3 | 0.0 |
| AutoRefine (Base) | 2.29 | 1.81 | 28.9 | 44.8 | 60.4 | 67.8 | 13.3 | 9.3 | 0.0 |
| AutoRefine, Base (+ exclude-seen) | 2.30 | 1.80 | 29.4 | 37.0 | 58.5 | 66.8 | 14.0 | 10.0 | 0.0 |
| R1-Searcher | 2.62 | 2.30 | 27.0 | 41.7 | 53.5 | 59.8 | 17.5 | 10.8 | 0.3 |
| R1-Searcher (+ exclude-seen) | 2.66 | 2.31 | 29.0 | 26.7 | 47.7 | 60.5 | 19.8 | 11.5 | 0.2 |
| AWA-RL (+ exclude-seen) | 3.33 | 3.04 | 20.2 | 34.7 | 46.6 | 53.6 | 20.9 | 5.8 | 0.0 |
| Qwen2.5-7B (no defense) | 1.28 | 1.10 | 15.1 | 46.4 | 56.6 | 60.6 | 5.6 | 4.7 | 50.5 |
| Qwen2.5-7B + TrustRAG | 1.29 | 1.28 | 15.0 | 42.3 | 47.7 | 33.4 | 7.5 | 7.3 | 53.0 |
| Qwen2.5-7B + Exact-match RL (1 seed) | 2.70 | 2.05 | 24.3 | 40.7 | 58.2 | 68.0 | 12.6 | 7.1 | 1.7 |
| Qwen2.5-7B + EchoSpec (+ xs, 3 seeds) | 3.07 | 3.05 | 23.5 | 12.2 | 20.4 | 25.1 | 22.5 | 14.4 | 34.3 |
| Qwen2.5-7B + EchoSpec (w/o xs, 1 seed) | 2.94 | 2.99 | 22.9 | 20.3 | 26.7 | 29.9 | 16.4 | 12.1 | 35.6 |
| Llama-3.1-8B (no defense) | 3.12 | 2.17 | 14.3 | 37.1 | 56.2 | 63.4 | 9.8 | 4.9 | 6.3 |
| Llama-3.1-8B + EchoSpec (+ xs, 1 seed) | 2.88 | 2.97 | 25.2 | 7.9 | 17.1 | 24.0 | 24.4 | 15.5 | 60.1 |
Table 11. Zero-shot cross-retriever transfer on Test Sets A and B without retraining: checkpoints trained only with e5-base-v2 are evaluated directly over 21M wiki-18 passages indexed by an unseen neural bi-encoder BAAI/bge-base-en-v1.5 [42] and by lexical Okapi BM25 [28], marking 1st and 2nd within each retriever block.
Abstention fairness and selective risk-coverage calibration. Under native prompts in Table 3, released RL agents stay at ≤ 0.2% Delete abstention. With an abstention prompt, Search-R1 reaches 36.6% versus 30.2% intact and Exact-match RL drops to 1.4–2.8%, whereas EchoSpec abstains on and 52.5% versus 27.8% and 25.0% intact. Across seven QA sets in Appendix § 11, it reaches 47.1% and 55.4% precision at 44.2 RA-F1, beating baselines by 2.8–3.0 points when answering while they miss 78–80% of refusals.
Ablations and behavioral mechanism
Table 4 separates primary structural drivers from targeted regularizers. Stage 1, Stage 2, and drive core gains: dropping Stage 1 leaves 9.4% Reach@8 and 69.4% Def@8, dropping Stage 2 leaves 47.2% Def@8, and cuts abstention to 20.6% and precision to 42.3%. With 32B warm-start, 1-seed and stay within seed noise on Inject(8) and inherit seed-1’s 39.6–40.7% abstention (49.7–50.0% Prec.), yet each guards its target regime: cuts Consistent(8) reach by 5.0 points (31.3% to 26.3%), while cuts self-guided 7B Def@8 from 56.0% to 51.5% and lifts Reach@8 from 20.5% to 23.5%. Disabling exclude-seen costs 3.3 points of Reach@8, and gives 12.8% Def@1 and 27.4% Reach@8 at 31.5% Def@8, favoring . Reach@8 tracks Def@8 in Figure 5a, and -search runs rise from 49% to 61% in Figure 5b.
Variant | Def@1 ↓ | Def@8 ↓ | Reach@8 ↑ | Con8-R ↑ | Acc@8 ↑ | Abst. ↑ | Prec. ↑ | RA-F1 ↑ EchoSpec (default, , 3 seeds) | 14.3 | 28.9 | 25.8 | 31.3 | 16.7 | 37.6 | 47.1 | 40.1 w/o Stage 2 RL (warm-start only, 3 seeds) | 20.8 | 47.2 | 21.0 | 23.4 | 13.9 | 33.1 | 47.0 | 41.5 w/o Stage 1 warm-start (-RL only, 1 seed) | 45.4 | 69.4 | 9.4 | 9.2 | 5.2 | 32.5 | 42.4 | 37.6 w/o wrong-answer penalty (, 1 seed) | 12.9 | 29.3 | 28.7 | 32.2 | 19.5 | 20.6 | 42.3 | 39.6 w/o evidence gate (, 1 seed) | 16.8 | 31.1 | 26.5 | 26.3 | 16.9 | 39.6 | 50.0 | 41.6 w/o acq. bonus (, 1 seed) | 16.2 | 29.1 | 26.2 | 30.1 | 15.8 | 40.7 | 49.7 | 41.7 w/o exclude-seen retrieval (3 seeds) | 17.0 | 31.6 | 22.5 | 25.8 | 15.2 | 40.3 | 47.9 | 40.1
cross-world baseline (, 3 seeds) | 12.8 | 31.5 | 27.4 | 29.9 | 16.1 | 31.6 | 43.3 | 37.6 w/o 32B teacher (self-guided 7B, 2 seeds) | 19.8 | 51.5 | 23.5 | 27.7 | 11.4 | 36.9 | 42.9 | 36.8 w/o 32B teacher & acq. bonus (3 seeds) | 21.7 | 56.0 | 20.5 | 24.7 | 9.0 | 43.1 | 44.1 | 37.2
Table 4. Table 4. Component ablations vs. shaded 3-seed default EchoSpec (±4.5 Def@8, ±1.5 Reach@8, ±3.3 Abst., ±1.2 Prec.). Con8-R: evidence reach on Consistent(8); 1-seed rows use seed 1 (32.4 Def@8, 39.6 Abst., 48.4 Prec.).
Figure 5. Figure 5. Behavioral mechanism. (a) Def@8 vs. Reach@8 across agents, backbones, and filters. (b) Searches per question across echo count over 3 seeds, where baselines stop early while EchoSpec searches deeper.
Conclusion and Limitations
Reinforcement-learned search agents follow paraphrased echoes because standard QA corpora conflate claim frequency with provenance, and they lose abstention because binary exact-match rewards leave wrong guesses unpriced. Larger budgets, forced turns, or passage filtering fail once policies stop on bare claims or query decoys. EchoSpec resolves both via five corpus transformations, a bridge-entity verification warm-start, and an evidence-gated -reward with an acquisition bonus and scale-preserving advantages, generalizing across Qwen2.5 and Llama-3.1 backbones and unseen dense and lexical BM25 retrievers.
Limitations. While we test up to eight turns over dense and BM25 indices of 21M Wikipedia passages, live web search adds dynamic rankings, SEO manipulation, and compromised sources, and world construction requires multi-hop items with known evidence, motivating open-ended multi-source synthesis.
Experimental Protocol and World Construction
Corpus and dense retrieval server. All training and default evaluation runs query the 21M-passage English Wikipedia dump wiki-18, segmented into 100-word chunks and indexed with e5-base-v2 [15, 20, 37]. Embeddings reside in fp16 on a dedicated GPU behind a micro-batched HTTP endpoint that supports per-request document overlays for Dup, Inject, Consistent, and Rewrite worlds, document-ID masks for Delete and Rewrite worlds, and per-episode exclude-seen filtering that masks passage IDs already returned earlier in the trajectory. Each query returns passages under a budget of four searches per episode. Across the seven standard QA benchmarks, this pipeline reproduces the released Search-R1-7B PPO checkpoint at 38.7 macro-averaged exact match versus the 38.5 originally reported by Jin et al. [15].
Symmetric world construction and retrieval-verified deletion. Base items come from the HotpotQA [48] and MuSiQue [34] development splits for evaluation and training splits for training, each paired with gold answer , bridge articles , and primary passage stating . For each item, an LLM generator produces diverse paraphrases of for Dup, a type-matched plausible decoy answer alongside paraphrased bare-claim passages stating without supporting derivation, and a minimally edited version of substituting for in Rewrite. Both Inject and Consistent reuse the same generated claim passages with only the answer string swapped between and across both title and body, eliminating stylistic, length, and perplexity cues between true and false echoes. For Delete, we iteratively probe top-30 e5 results using , the primary title with , and for up to four rounds, masking answer-bearing passages related to until a probe round surfaces zero leaks, and discard unverified items. Two generators from distinct model families build disjoint ten-world evaluation suites spanning Identity, Dup, Inject, Consistent, Rewrite, and Delete: Qwen3-32B [46] forms Test Set A with 507 items for 5,070 episodes, and Phi-4 [1] forms Test Set B with 607 items for 6,070 episodes. A held-out TF-IDF classifier reaches only 0.60–0.68 AUC separating true from false echo claims across both test suites.
Training, Warm-Start Synthesis, and Hyperparameters
Warm-start trajectory synthesis and filtering. Training worlds cover 3,968 multi-hop questions from the HotpotQA and MuSiQue training splits with Qwen3-32B claim texts, sharing zero questions with Test Set A and neither questions nor generator family with Phi-4 Test Set B. For Stage 1 synthesis in Table 5, we sample three worlds per item with probabilities 0.60 on Inject, 0.15 on Delete, 0.15 on Dup, and 0.10 on Consistent, weighting echo counts at 0.10, 0.20, 0.30, 0.40, and roll out four trajectories per world under temporary article-chain guidance. Filtering retains 3,582 multi-hop trajectories under Qwen2.5-32B-Instruct guidance at 2.25 mean searches, comprising 940 Inject with 541 at , 246 Delete, 447 Dup, 225 Consistent, and 1,724 identity trajectories. Self-guided Qwen2.5-7B-Instruct synthesis retains 2,537 trajectories at 2.13 mean searches across 773 Inject, 308 Delete, 267 Rewrite, 304 Dup, 187 Consistent, and 698 identity episodes. Across 39,568 candidate 7B rollouts, 7,897 fail hop continuity, 3,930 fail grounded outcome or depth, and 1,213 mention the guidance note, leaving self-contained multi-hop demonstrations once the temporary guidance note is removed before supervised fine-tuning.
| Stage 1 Parameter | Value | Stage 2 Parameter | Value |
| Base backbone | Qwen2.5-7B-Instruct | Rollouts per world | 6 identity + 6 transform |
| LoRA rank / | 64 / 128 on linear layers | Questions per step | 12, giving 144 episodes |
| Worlds per train item | 1 identity + 2 transforms | Policy-gradient steps | 200, lr , warmup 10 |
| Synthesis rollouts / temp | 4 per world / | KL penalty / grad clip | / 1.0 |
| Retained worlds, 32B / 7B | 3,582 / 2,537 of 9,892 | Reward | (0.10, 0.25, 0.30) |
| SFT updates / batch / lr | 60 steps / 64 / | Cross-world baseline | 0 default or 0.35 |
Table 5. Hyperparameters for EchoSpec Stage 1 warm-start synthesis and SFT alongside Stage 2 -calibrated group RL.
Optimization and compute. Stage 1 fine-tunes LoRA [14] adapters at rank , , and context limit 6,144 while masking prompt and <information> tokens for 60 AdamW steps. Stage 2 runs 200 asynchronous GRPO steps with policy lag at most two updates over world weights 0.10 Dup, 0.40 Inject, 0.15 Consistent, 0.10 Rewrite, and 0.25 Delete, and echo weights 0.10, 0.10, 0.20, 0.25, 0.35 on . Each full two-stage training run completes in roughly two wall-clock hours on four 192 GB AMD Instinct MI300X GPUs sharing one dedicated GPU-resident retrieval server.
Extended Theoretical Analysis
Proposition 3 (Terminal threshold invariance under outcome-independent acquisition bonuses). Consider episode return at a terminal state with retrieved passages , where is any bounded function of alone rather than a step-wise potential difference . Let denote the conditional correctness probability of candidate answer on an answerable world, with if and if . Answering beats ABSTAIN in expected return if and only if , invariant to both the weight and the choice of terminal bonus .
Proof. Conditioned on fixed retrieved set , choosing yields expected return , whereas choosing ABSTAIN yields . Subtracting cancels identically for every , giving , which characterizes the reward’s Bayes-optimal stopping rule given rather than calibrated posteriors under finite-sample RL. When so , this recovers , whereas when so , the threshold rises to , favoring further search or abstention until primary passage enters . □
Proposition 4 (Unbiasedness and tie-breaking of the cross-world control variate). For any question and invariant transformation , let and denote returns from independent rollouts in and , and set with for . Replacing with adds zero bias to the policy gradient in up to standard leave-one-out scaling, and whenever all rollouts in defer to echoes with , every echo-deferring rollout receives strictly negative advantage across the group.
Proof. Because are sampled independently of trajectory given and , the cross-world baseline satisfies . When , the within-world advantage vanishes identically, whereas , penalizing uniform echo deference on questions that the policy already answers cleanly on . □
Per-Generator and Per-World Evaluation Breakdown
Table 6 disaggregates all ten transformation worlds separately on Qwen3-32B Test Set A and unseen Phi-4 Test Set B. First, EchoSpec transfers substantially to unseen Phi-4 echoes on Test Set B: standalone Qwen2.5-7B EchoSpec achieves 19.4%, 33.8%, and 38.8% conditional deference, 24.0% Reach@8, 15.4% Acc@8, and 37.6% Delete abstention, while Llama-3.1-8B EchoSpec reaches 16.4%, 33.9%, and 37.7% deference, 24.6% Reach@8, 16.0% Acc@8, and 52.7% abstention. This cuts Def@8 by 21.7–22.8 points below Qwen2.5-7B at 60.5% and by 18.3–29.9 points below every released RL agent at 57.1–67.6% while more than tripling Reach@8 from 7.7% to 24.0–24.6%, and pairing Qwen2.5-7B EchoSpec with TrustRAG reaches 15.5% Phi-4 Def@8. Second, comparing the two suites reveals a cross-generator gap in standalone Def@8 between 19.0–23.6% on Test Set A and 37.7–38.8% on Test Set B even though Reach@8 transfers closely from 27.7–28.4% on A to 24.0–24.6% on B. Conditioning on retrieval pinpoints why: once primary passage is retrieved, Def@8 stays low on both generators at 5.6% on A versus 13.0% on B for Qwen2.5-7B EchoSpec and 11.5% versus 14.5% for Llama-3.1-8B EchoSpec. Nearly the entire cross-generator shift occurs when is missed, where Def@8 rises from 31.9% at 30.6% abstention on A to 57.0% at 15.2% abstention on B for Qwen2.5-7B and from 38.7% to 54.2% for Llama-3.1-8B, showing that both evidence acquisition and post-retrieval verification generalize cleanly across generators whereas fallback abstention on unretrieved items still benefits from familiarity with training-generator phrasing. On Rewrite worlds where itself is edited to state , EchoSpec reaches 13.5–13.6% and 16.9–18.1% accuracy on A and B, nearly doubling Qwen2.5-7B at 7.9% and 10.2% and beating -reward RL at 12.2% and 15.0%, while trailing non-abstaining AutoRefine at 20.1% and 23.1% because EchoSpec occasionally abstains when edited passages conflict with parametric priors. On Dup(3), TrustRAG lowers base accuracy from 16.8% and 21.1% to 15.0% and 16.0% by clustering away valid supporting copies, whereas EchoSpec reaches 29.1–32.4% and 31.9–34.9%. Across single-stage RL runs varying , Delete abstention rises across 31.3%, 40.6%, 42.2%, and 44.9% as clean accuracy trades off from 24.7% down to 20.9% at higher values.
| Method | Id. ↑ | Dup3 ↑ | Con8 ↑ | Rew. ↑ | Def@1 ↓ | Def@3 ↓ | Def@8 ↓ | Reach@8 ↑ | Acc@8 ↑ | Abst. ↑ |
| Test Set A: Qwen3-32B generator, 507 questions × 10 worlds | ||||||||||
| Search-R1 | 24.1 | 28.2 | 62.5 | 15.0 | 47.5 | 63.1 | 70.5 | 11.1 | 5.5 | 0.0 |
| AutoRefine (Base) | 32.1 | 35.9 | 68.6 | 20.1 | 41.1 | 56.4 | 62.6 | 16.6 | 10.7 | 0.0 |
| R1-Searcher | 31.4 | 34.9 | 66.7 | 19.5 | 35.2 | 55.4 | 64.8 | 20.7 | 11.1 | 0.0 |
| AWA-RL | 23.5 | 24.5 | 47.1 | 12.2 | 52.9 | 52.9 | 57.1 | 19.7 | 4.1 | 0.0 |
| Qwen2.5-7B (no defense) | 16.6 | 16.8 | 51.7 | 7.9 | 38.1 | 58.3 | 57.1 | 6.5 | 4.3 | 50.7 |
| Qwen2.5-7B + TrustRAG | 15.4 | 15.0 | 25.1 | 7.3 | 37.2 | 41.0 | 20.5 | 12.0 | 9.1 | 48.1 |
| -reward RL | 19.7 | 22.9 | 61.1 | 12.2 | 50.0 | 69.0 | 71.0 | 8.7 | 4.9 | 33.1 |
| Qwen2.5-7B + EchoSpec | 25.9 | 29.1 | 44.2 | 13.5 | 9.1 | 14.0 | 19.0 | 27.7 | 17.9 | 37.5 |
| Llama-3.1-8B + EchoSpec | 27.6 | 32.4 | 44.6 | 13.6 | 5.7 | 14.3 | 23.6 | 28.4 | 18.3 | 52.3 |
| Qwen2.5-7B + EchoSpec + TrustRAG | 25.4 | 23.4 | 28.3 | 13.7 | 10.3 | 12.4 | 10.1 | 31.8 | 19.9 | 37.2 |
| Qwen2.5-7B + EchoSpec + Sem. dedup | 25.3 | 28.8 | 47.3 | 13.1 | 9.5 | 19.0 | 22.1 | 39.1 | 22.0 | 33.7 |
| Test Set B: Phi-4 generator, unseen family, 607 questions × 10 worlds | ||||||||||
| Search-R1 | 28.5 | 32.6 | 50.7 | 19.3 | 44.5 | 63.6 | 67.6 | 15.5 | 8.1 | 0.3 |
| AutoRefine (Base) | 33.8 | 37.7 | 57.2 | 23.1 | 40.0 | 58.0 | 57.1 | 19.6 | 14.3 | 0.0 |
| R1-Searcher | 32.6 | 37.7 | 57.5 | 21.6 | 31.3 | 53.0 | 59.1 | 21.2 | 14.0 | 0.0 |
| AWA-RL | 25.2 | 30.3 | 47.5 | 16.0 | 43.8 | 58.2 | 60.1 | 19.9 | 7.9 | 0.0 |
| Qwen2.5-7B (no defense) | 18.8 | 21.1 | 42.5 | 10.2 | 45.6 | 56.1 | 60.5 | 7.7 | 6.8 | 47.8 |
| Qwen2.5-7B + TrustRAG | 17.1 | 16.0 | 20.8 | 9.6 | 33.6 | 42.3 | 25.0 | 10.7 | 9.7 | 47.8 |
| -reward RL | 22.6 | 27.4 | 50.1 | 15.0 | 40.9 | 67.9 | 67.9 | 10.2 | 5.4 | 32.0 |
| Qwen2.5-7B + EchoSpec | 28.6 | 31.9 | 43.3 | 16.9 | 19.4 | 33.8 | 38.8 | 24.0 | 15.4 | 37.6 |
| Llama-3.1-8B + EchoSpec | 30.1 | 34.9 | 45.1 | 18.1 | 16.4 | 33.9 | 37.7 | 24.6 | 16.0 | 52.7 |
| Qwen2.5-7B + EchoSpec + TrustRAG | 27.3 | 27.0 | 27.6 | 16.9 | 17.4 | 26.5 | 15.5 | 27.1 | 18.2 | 35.7 |
| Qwen2.5-7B + EchoSpec + Sem. dedup | 28.2 | 30.8 | 42.4 | 17.3 | 23.8 | 32.8 | 35.2 | 33.3 | 18.3 | 34.1 |
Table 6. Per-generator and per-world breakdown separating Qwen3-32B Test Set A on disjoint questions from Phi-4 Test Set B. Columns give clean accuracy Id., duplicated-evidence accuracy Dup3, 8-true-echo accuracy Con8, rewritten-evidence accuracy Rew., conditional deference at , primary-evidence reach and EM at , and Delete abstention, marking standalone 1st and 2nd plus shaded filter compositions beating standalone best.
Per-Benchmark Results on Standard QA Datasets
Table 7 reports per-benchmark answering precision alongside macro-averaged exact match and reliability-aware F1 across the three single-hop benchmarks NQ [22], TriviaQA [16], and PopQA [24] and the four multi-hop benchmarks HotpotQA [48], 2WikiMultiHopQA [13], MuSiQue [34], and Bamboogle [27]. Released outcome-reward RL agents never abstain, equating their precision with raw exact match. Base AutoRefine attains the highest raw exact match and RA-F1 at 44.7%, whereas multi-hop datasets with incomplete retrieval yield 11.2–22.7% precision on MuSiQue when agents always guess. Conversely, MADAM-RAG reaches 55.7% precision only by abstaining on 66.0% of standard QA questions, collapsing its macro-averaged exact match to 18.9% and RA-F1 to 28.3%. Relative to Search-R1 at 38.4% Avg EM and 38.4% Avg Prec.,
| Method | NQ | TriviaQA | PopQA | HotpotQA | 2Wiki | MuSiQue | Bamboogle | Avg EM | Avg Prec. | RA-F1 |
| Released search agents | ||||||||||
| Search-R1 | 39.9 | 60.7 | 40.0 | 38.8 | 33.1 | 17.2 | 39.2 | 38.4 | 38.4 | 38.4 |
| ZeroSearch | 38.8 | 61.7 | 37.9 | 33.3 | 32.0 | 11.2 | 31.2 | 35.1 | 35.2 | 35.2 |
| AutoRefine (Instruct) | 43.5 | 62.8 | 42.2 | 42.5 | 40.2 | 21.6 | 44.0 | 42.4 | 42.4 | 42.4 |
| AutoRefine (Base) | 47.7 | 64.8 | 45.6 | 46.5 | 38.9 | 22.5 | 47.2 | 44.7 | 44.7 | 44.7 |
| ReSearch | 38.7 | 61.9 | 42.7 | 42.6 | 36.1 | 22.7 | 39.2 | 40.6 | 40.6 | 40.6 |
| R1-Searcher | 38.4 | 59.2 | 35.6 | 45.5 | 45.8 | 22.2 | 45.6 | 41.7 | 41.8 | 41.7 |
| AWA-RL | 31.3 | 55.2 | 40.2 | 37.8 | 27.7 | 15.9 | 43.2 | 35.8 | 35.9 | 35.9 |
| Qwen2.5-7B with inference-time defenses | ||||||||||
| No defense | 39.0 | 65.1 | 44.7 | 40.3 | 33.2 | 18.8 | 36.4 | 28.8 | 41.1 | 33.8 |
| TrustRAG | 37.7 | 63.4 | 43.7 | 39.8 | 32.9 | 18.9 | 35.5 | 28.0 | 40.3 | 33.0 |
| MADAM-RAG† | 46.3 | 74.5 | 58.9 | 51.4 | 32.3 | 28.9 | 47.6 | 18.9 | 55.7 | 28.3 |
| Qwen2.5-7B trained on the same questions | ||||||||||
| Exact-match RL | 37.2 | 60.6 | 36.8 | 38.0 | 31.6 | 19.6 | 36.0 | 36.5 | 37.2 | 36.8 |
| λ-reward RL | 37.8 | 63.9 | 45.1 | 41.4 | 37.0 | 20.3 | 45.6 | 33.8 | 42.4 | 37.6 |
| EchoSpec (3 seeds) | 42.5 | 66.7 | 47.6 | 49.0 | 42.5 | 28.5 | 48.0 | 34.9 | 47.1 | 40.1 |
| Llama-3.1-8B cross-family replication (1 seed) | ||||||||||
| No defense | 29.4 | 51.5 | 35.6 | 26.7 | 18.1 | 10.5 | 26.2 | 26.8 | 28.4 | 27.6 |
| Exact-match RL | 41.7 | 64.5 | 40.5 | 39.9 | 41.3 | 21.6 | 48.0 | 41.7 | 42.7 | 42.2 |
| EchoSpec | 48.2 | 71.6 | 55.6 | 56.5 | 56.6 | 36.7 | 56.7 | 36.7 | 55.4 | 44.2 |
| Qwen2.5-7B with orthogonal retrieval filters (3 seeds) | ||||||||||
| EchoSpec + TrustRAG | 40.4 | 65.2 | 46.6 | 47.1 | 36.2 | 26.3 | 42.5 | 34.0 | 44.1 | 38.3 |
| EchoSpec + Sem. dedup | 42.2 | 66.1 | 49.2 | 46.9 | 40.1 | 26.2 | 47.4 | 35.3 | 46.2 | 40.0 |
Table 7. Answering precision (%) across the seven standard QA benchmarks, with macro Exact Match Avg EM, macro Precision Avg Prec., and Reliability-Aware F1 RA-F1. † MADAM-RAG abstains on 66.0% of QA items at 18.9% Avg EM, marking standalone 1st and 2nd plus shaded filter compositions beating standalone best.
Qwen2.5-7B EchoSpec trades 3.5 points of exact-match coverage at 34.9% Avg EM to raise answering precision on every benchmark: +2.6 points on NQ, +6.0 on TriviaQA, +7.6 on PopQA, +10.2 on HotpotQA, +9.4 on 2WikiMultiHopQA, +11.3 on MuSiQue, and +8.8 on Bamboogle, lifting macro precision to 47.1% and RA-F1 to 40.1. On Llama-3.1-8B (1 seed), EchoSpec attains the highest standalone precision on all seven datasets at 48.2–71.6%, lifting macro precision from 28.4% base and 42.7% Exact-match RL to 55.4% at 36.7% Avg EM and 44.2 RA-F1. Partitioning each benchmark by EchoSpec’s answering decision rules out easy-item selection by abstention: on the 66.3% of items where Llama-3.1-8B EchoSpec answers, it outperforms Exact-match RL on the exact same items at 54.6% versus 51.6% accuracy, whereas on the 33.7% of items where EchoSpec abstains, Exact-match RL errs on 80.4% of questions at only 19.6% accuracy. The same selective-calibration pattern holds on Qwen2.5-7B at 74.2% coverage, beating Search-R1 on answered items at 46.4% versus 43.6% while Search-R1 errs on 77.9% of items EchoSpec refuses.
Fairness Controls, Failure Modes, and Reproducibility
Abstention under deleted evidence and updating under rewritten evidence. Beyond multi-hop recovery under echoes, Appendix § 14 provides full turn-by-turn benchmark trajectories, while two complementary behaviors govern Delete and Rewrite worlds. When primary evidence bearing Colin Kaepernick’s preferred childhood nickname is masked from wiki-18, Search-R1 repeats variants of Colin Kaepernick preferred nickname three times and hedges in free text without abstaining, which scores as a wrong answer, while Exact-match RL fabricates a plausible nickname. EchoSpec queries Colin Kaepernick, Colin Kaepernick biography, and Colin Kaepernick nickname across distinct unseen biographical passages, verifies that none supports the requested fact, and outputs ABSTAIN. Conversely, on Rewrite episodes where the primary passage itself is edited to state , EchoSpec locates the edited primary passage and reports rather than falling back on stale parametric memory, separating verification from rigid refusal.
Unified exclude-seen retrieval fairness. Because released RL checkpoints in Table 3 run in their native evaluation loop without episode-level passage masking while default EchoSpec enables exclude-seen retrieval, Table 8 tests both directions symmetrically across all nine suites. First, granting exclude-seen retrieval to all seven released RL search agents and all five inference defenses barely alters baseline vulnerability: released RL agents remain at 59.7–74.7% Def@8, compared with 58.6–72.9% under standard retrieval, because outcome-trained agents halt after one or two queries as soon as an echo states a candidate answer rather than issuing follow-up entity queries that benefit from unseen-passage masking. Second, disabling exclude-seen on the 3-seed Qwen2.5-7B EchoSpec checkpoints raises Def@8 by only 2.7 points from 28.9% to 31.6%, and disabling exclude-seen throughout both training and evaluation yields 27.1% Def@8 while lowering Reach@8 from 25.8% to 23.6% and Acc@8 from 16.7% to 14.3%. Exclude-seen retrieval therefore acts as a slot-efficiency foil for policies that already reformulate queries past echoes, whereas Qwen2.5-7B EchoSpec beats all seven released RL agents by 30–44 points under standard baselines and reaches 28.9% versus 59.7–74.7% when every baseline is also granted episode-level exclude-seen retrieval.
Prompt protocols and abstention fairness. In Table 3, external checkpoints run in their native prompt and tag formats, using <search> and <answer> for Search-R1 [15], ZeroSearch [33], AutoRefine [30], ReSearch [4], and R1-Searcher [31], and <tool_call> with JSON <answer> and special-token retention for AWA-RL [49]. Our abstention detector matches both ABSTAIN and natural-language refusals such as “I do not know”, “insufficient information”, “cannot be determined”, and “unable to answer”. Because native outcome-RL prompts omit an explicit refusal token, we run two fairness controls. First, evaluating released Search-R1 under our standard prompt permitting abstention yields 36.6% DELETE abstention versus 30.2% on intact evidence, or 34.7% with exclude-seen in Table 8, with 34.3% EM, 44.0% Prec., and 38.5% RA-F1 across the seven QA sets, confirming weaker sensitivity to missing evidence than Qwen2.5-7B at 49.2% and Llama-3.1-8B EchoSpec at 52.5% versus 25.0% intact. Second, training Exact-match RL directly under that same prompt still drives DELETE abstention down to 1.4% on Qwen2.5-7B and 2.8% on Llama-3.1-8B. For AWA-RL, auditing all Test Set A DELETE episodes shows 0.0% explicit refusals versus 0.2–1.1% on standard QA sets, 26.8% wrong guesses, 13.6% per-turn length truncations, and 59.6% episodes requesting a fifth <tool_call> at , while doubling AWA-RL’s search budget to in Table 10 still yields only
0.5–0.6% Delete abstention and 61.8–67.2% Def@8. Semantic deduplication in § 5.1 applies greedy cosine filtering at in e5-base-v2 space before top-3 truncation without source labels.
Reproducibility. World-construction pipelines in echospec/transforms.py, trajectory synthesis and three-gate filtering in echospec/synth.py, distributed LoRA SFT and asynchronous GRPO training in echospec/train.py, dense and BM25 retrieval and evaluation modules in echospec/retrieval.py and echospec/evaluate.py, shell workflows in scripts/, and reference parameters and split manifests in configs/ are all publicly released in our open-source repository for full reproduction.
Cross-Backbone, Budget, Forced-Turn, and Cross-Retriever Evaluation
Replication across model scales and families. Table 9 evaluates echo deference and EchoSpec training across Qwen2.5-3B-Instruct, Qwen2.5-7B-Instruct [47], and Llama-3.1-8B-Instruct [11]. Without defense, Qwen2.5-3B defers at 62.5% under eight echoes with 5.1% Reach@8 and 2.8% Acc@8, while Llama-3.1-8B defers at 70.0% and 74.4% under exclude-seen with 10.8% Reach@8 and 5.0% Delete abstention. Applying Exact-match RL amplifies deference to 71.5% on Qwen2.5-3B and keeps it at 66.7% on Llama-3.1-8B while eroding Delete abstention to 5.2% and 2.8%. In contrast, EchoSpec converges all three backbones to 28.9–30.6% Def@8 without hyperparameter retuning: Llama-3.1-8B EchoSpec reaches 11.1 / 24.1 / 30.6% deference, 26.5% Reach@8, 17.2% Acc@8, 52.5% abstention, and 55.4% QA precision at 44.2 RA-F1, while Qwen2.5-3B EchoSpec reaches 15.3 / 29.6 / 30.4% deference, 22.5% Reach@8, and 42.8% precision, outperforming all seven 7B RL baselines in deference and reach at less than half the parameter count.
Search budget scaling and forced multi-turn verification. Table 10 evaluates both enlarged search caps and across all three EchoSpec seeds alongside protocol-enforced multi-turn verification baselines that intercept early answers before three searches under exclude-seen retrieval. Even at with exclude-seen, Search-R1 and Base AutoRefine stop voluntarily after 2.03 and 1.91 searches under eight echoes and defer at 67.4% and 59.8%, while AWA-RL issues 3.65 searches yet defers at 61.8%. Even when Search-R1 and Qwen2.5-7B are forced under an explicit bridge-verification prompt to execute at least three distinct searches, Srch@8 rises to 3.18 and 3.01 and Reach@8 climbs to 29.5% and 27.5%, yet Def@8 remains high at 55.5–58.7% and 54.1% with only 9.8–11.0% and 7.8% Acc@8 because untrained policies issue confirmatory decoy queries and still defer to eight-echo majorities after retrieving . By contrast, 3-seed EchoSpec achieves 28.9% Def@8 and 16.7% Acc@8 at , improves to 18.3% Acc@8 at , and reaches 34.8% Reach@8 and 20.2% Acc@8 at as wider retrieval surfaces primary passages earlier.
Zero-shot transfer across dense and lexical retrievers. To test whether EchoSpec learns retriever-agnostic verification rather than overfitting to e5-base-v2 rankings, Table 11 evaluates frozen e5-trained checkpoints directly under BAAI/bge-base-en-v1.5 [42] and lexical Okapi BM25 [28] over all 21M wiki-18 passages without retraining. Under BGE, released RL agents and Exact-match RL remain at 58.1–72.2% Def@8, whereas 3-seed Qwen2.5-7B EchoSpec achieves 13.4 / 21.2 / 29.4% deference with 26.0% Reach@8 and 15.2% Acc@8, matching in-retriever e5 within 0.5 points, and Llama-3.1-8B EchoSpec reaches 11.3 / 20.8 / 30.5% deference with 25.2% Reach@8 and 15.4% Acc@8. Under non-neural Okapi BM25, where TrustRAG degrades from 22.8% to 33.4% Def@8 at 7.5% Reach@8 and RL baselines stay at 53.6–68.0% Def@8, standalone EchoSpec achieves the lowest deference overall at 12.2 / 20.4 / 25.1% across 3 Qwen2.5-7B seeds, or 29.9% even without exclude-seen, and 7.9 / 17.1 / 24.0% on Llama-3.1-8B while quadrupling base Reach@8 from 5.6% to 22.5–24.4% across both generator-disjoint test suites.
Step-by-Step Trajectory Examples on Real Benchmarks
To illustrate how EchoSpec alters multi-turn search behavior in practice, Table 12 to Table 15 present verbatim evaluation trajectories from our logs comparing released RL search agents and inference-time TrustRAG against EchoSpec. Table 12 and Table 13 come from unseen Phi-4 eight-echo Inject worlds on HotpotQA and MuSiQue in Test Set B, while Table 14 and Table 15 come from unmodified clean HotpotQA and 2WikiMulti-HopQA over standard wiki-18 without any synthetic transformation. Across both settings, outcome-RL baselines either stop as soon as a Turn-1 passage states a plausible answer or latch onto homonym and distractor entities without verifying intermediate hops, whereas EchoSpec reformulates entity-focused queries under exclude-seen retrieval until confirming the target answer directly in the primary Wikipedia article.
| Question: The person that William Flannery was an architect for entered the publishing business in what year? Gold: 1887 Echo ×8: 1890 | |||
| Agent | Search trajectory and retrieved evidence summary | Answer | Result |
| Search-R1 | Turn 1: <search>William Flannery architect for</search> → top-3 results are Phi-4 echoes stating “William Flannery, a renowned architect, worked for William Randolph Hearst, who entered the publishing business in 1890...” Stops immediately without querying William Randolph Hearst. | 1890 | ✗ Deferred |
| AutoRefine | Turn 1: <search>William Flannery architect for</search> → retrieves the same three bare echo claims asserting 1890 and terminates after one search. | 1890 | ✗ Deferred |
| TrustRAG | Turn 1: <search>William Flannery architect client publishing business entry year</search> → clustering removes the injected echoes from context, leaving no primary passage on Hearst’s publishing entry, and the base model halts. | ABSTAIN | ✗ Missed |
| EchoSpec | Turn 1: <search>William Flannery architect</search> → surfaces echoes alongside bridge article William Flannery reading “American art director and architect for William Randolph Hearst.” Refuses to answer from unverified bare claims. Turn 2: <search>William Randolph Hearst biography</search> → targets the resolved patron entity directly under exclude-seen retrieval. Turn 3: <search>William Randolph Hearst entered publishing business</search> → retrieves primary article William Randolph Hearst reading “Hearst entered the publishing business in 1887 after being given control of The San Francisco Examiner by his wealthy father.” | 1887 | ✓ Grounded |
Table 12. Trajectory comparison on HotpotQA under unseen Phi-4 echoes in Test Set B Inject at . Eight paraphrased Phi-4 claim pages state decoy year 1890, while primary Wikipedia articles William Flannery and William Randolph Hearst remain intact and retrievable throughout the standard wiki-18 corpus.
| Question: Who wrote the song sung by the performer of The Story? | Gold: Phil Hanseroth | Echo ×8: Tim Hanseroth | |
| Agent | Search trajectory and retrieved evidence summary | Answer | Result |
| AutoRefine | Turn 1: <search>"The Story" performer</search> → top results return Phi-4 echoes stating “Tim Hanseroth is the acclaimed songwriter behind the song ‘The Story,’ performed by Brandi Carlile.” Stops at Turn 1. | Tim Hanseroth | ✗ Deferred |
| Search-R1 | Turn 1: <search>who wrote the song sung by the performer of The Story</search> → retrieves Phi-4 echoes naming Tim Hanseroth. Turn 2: <search>Tim Hanseroth</search> → queries the echo candidate rather than the song, sees that Tim Hanseroth is indeed a songwriter for Brandi Carlile on Wikipedia, and answers Tim Hanseroth. | Tim Hanseroth | ✗ Deferred |
| TrustRAG | Turn 1: <search>"The Story" performer song writer</search> → clusters away echoes. Turns 2–3: <search>"Grace in Gravity" song writer</search>, <search>"Grace in Gravity" The Story writer</search> → drifts onto an unrelated album track and abstains. | ABSTAIN | ✗ Missed |
| EchoSpec | Turn 1: <search>Who performed The Story?</search> → echoes appear without primary album or single credits. Turn 2: <search>Performer of The Story song</search> → exclude-seen surfaces bridge article The Story (Brandi Carlile album) confirming performer Brandi Carlile. Turn 3: <search>Who wrote The Story song by Brandi Carlile?</search> → retrieves primary article The Story (song) reading “a song released as a single by American folk rock singer Brandi Carlile, written by Phil Hanseroth.” | Phil Hanseroth | ✓ Grounded |
Table 13. Trajectory comparison on MuSiQue under unseen Phi-4 echoes in Test Set B Inject at . Search-R1 issues a second confirmatory query containing decoy name Tim Hanseroth, Brandi Carlile’s twin bandmate, whereas EchoSpec resolves performer Brandi Carlile first and queries the primary song article directly.
Table 14. Trajectory comparison on clean unmodified HotpotQA over standard wiki-18 with no synthetic overlays. Released RL baselines stop on distractor mansion owners when Turn-1 dense retrieval misses the bridge entity, whereas EchoSpec reformulates entity-focused queries until grounding both intermediate hops in Wikipedia.
| Question: Indian Mound Cottage was owned by the financier that co-founded which company? | Gold: Standard Oil | Suite: HotpotQA | |
| Agent | Search trajectory and retrieved evidence summary | Answer | Result |
| Search-R1 | Turn 1: <search>Indian Mound Cottage financier co-founded company</search> → composite query misses Indian Mound Cottage in top-3 e5 results and surfaces a distractor passage on real-estate financiers. Stops after Turn 1. | American Realty Capital | ✗ Wrong |
| AutoRefine | Turns 1–2: <search>Indian Mound Cottage owner</search>, <search>Indian Mound Cottage financier owner</search> → picks up a distractor passage mentioning businessman Barry Schwartz without verifying ownership of Indian Mound Cottage.<br>Turn 3: <search>Barry Schwartz Calvin Klein co-founder</search> → answers Calvin Klein Inc. | Calvin Klein Inc. | ✗ Wrong |
| R1-Searcher | Turns 1–3: <search>Indian Mound Cottage owner</search>, <search>Ernest Otter co-founded which company</search>, <search>Ernest Otter owned Indian Mound Cottage</search> → chases an unverified entity and guesses Guinness plc. | Guinness plc | ✗ Wrong |
| EchoSpec | Turns 1–2: <search>Indian Mound Cottage owner</search>, <search>owner of Indian Mound Cottage</search> → rejects unverified distractors while exclude-seen clears top ranks.<br>Turn 3: <search>financier who co-founded a company and owned Indian Mound Cottage</search> → retrieves bridge article Indian Mound Cottage reading “mansion that was owned by oil executive William Rockefeller.”<br>Turn 4: <search>William Rockefeller Indian Mound Cottage</search> → retrieves primary article William Rockefeller reading “American businessman and financier... co-founder of Standard Oil.” | Standard Oil | ✓ Grounded |
caption:
Table 15. Trajectory comparison on clean unmodified 2WikiMultiHopQA over standard wiki-18. All agents identify songwriter Arthur Lee in Turn 1 search results, after which Turn 2 surfaces homonym Arthur Lee, 1st Viscount Lee of Fareham, and only EchoSpec disambiguates the rock musician entity before answering.
| Question: Where was the place of death of the composer of song A House Is Not A Motel? | Gold: Memphis, Tennessee | Suite: 2Wiki | |
| Agent | Search trajectory and retrieved evidence summary | Answer | Result |
| AutoRefine | Turn 1: <search>composer of the song A House Is Not A Motel</search> → identifies Arthur Lee of the rock band Love.<br>Turn 2: <search>place of death of Arthur Lee</search> → retrieves British politician Arthur Lee, 1st Viscount Lee of Fareham, who died at Avening, Gloucestershire, and answers without checking whether he wrote the song. | Avening, Gloucestershire | ✗ Wrong |
| R1-Searcher | Turn 1: <search>A House Is Not A Motel composer</search> → identifies Arthur Lee.<br>Turn 2: <search>Arthur Lee place of death</search> → accepts the same homonym biography and outputs Avening, Gloucestershire. | Avening, Gloucestershire | ✗ Wrong |
| Search-R1 | Turns 1–4: <search>composer of the song A House Is Not A Motel</search>, <search>place of death of Arthur Lee</search>, <search>Arthur Lee composer place of death</search>, <search>Arthur Lee composer of A House Is Not A Motel place of death</search> → repeats overlapping queries without exclude-seen or musician disambiguation and hits the four-search cap with no answer. | No answer | ✗ Cap hit |
| EchoSpec | Turn 1: <search>composer of song A House Is Not A Motel</search> → resolves songwriter Arthur Lee of Love.<br>Turn 2: <search>Arthur Lee death place</search> → sees non-musician Arthur Lee biography and rejects the mismatch.<br>Turn 3: <search>Arthur Lee musician death place</search> → disambiguates by profession and retrieves Arthur Lee (musician), confirming his death from leukemia in Memphis, Tennessee. | Memphis, Tennessee | ✓ Grounded |
caption:
References
- [1]Marah Abdin, Jyoti Aneja, Harkirat Behl, Sébastien Bubeck, Ronen Eldan, Suriya Gunasekar, Michael Harrison, Russell J. Hewett, Mojan Javaheripi, Piero Kauffmann, James R. Lee, Yin Tat Lee, Yuanzhi Li, Weishung Liu, Caio C. T. Mendes, Anh Nguyen, Eric Price, Gustavo de Rosa, Olli Saarikivi, Adil Salim, Shital Shah, Xin Wang, Rachel Ward, Yue Wu, Dingli Yu, Cyril Zhang, and Yi Zhang. Phi-4 technical report. arXiv preprint arXiv:2412.08905, 2024.arxiv.org/abs/2412.08905
- [2]Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019.arxiv.org/abs/1907.02893
- [3]Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannanen Hajishirzi. Self-RAG: Learning to retrieve, generate, and critique through self-reflection. In International Conference on Learning Representations, 2024.arxiv.org/abs/2310.11511
- [4]Mingyang Chen, Linzhuang Sun, Tianpeng Li, Haoze Sun, Yijie Zhou, Chenzheng Zhu, Haofen Wang, Jeff Z. Pan, Wen Zhang, Huajun Chen, Fan Yang, Zenan Zhou, and Weipeng Chen. ReSearch: Learning to reason with search for LLMs via reinforcement learning. In Advances in Neural Information Processing Systems, 2025.arxiv.org/abs/2503.19470
- [5]Yujia Chen, Yuhui Wang, Junzhe Wang, Shihan Dou, Di Liang, Yujiong Shen, Yuhao Zhou, Zhiheng Xi, Ming Zhang, Senjie Jin, Jiayi Chen, Dingwei Zhu, Yajie Yang, Yuan Hua, Chenyao Lu, Tao Gui, Qi Zhang, and Xuanjing Huang. QREF: Improving deepsearch agents by aligning search with reasoning via query refinement and evidence filtering. arXiv preprint arXiv:2610.21512, 2026.arxiv.org/abs/2610.21512
- [6]Ran El-Yaniv and Yair Wiener. On the foundations of noise-free selective classification. Journal of Machine Learning Research, 11:1605–1641, 2010.DOI
- [7]Lisa K. Fazio, Nadia M. Brashier, B. Keith Payne, and Elizabeth J. Marsh. Knowledge does not protect against illusory truth. Journal of Experimental Psychology: General, 144(5):993–1002, 2015.DOI
- [8]Shangbin Feng, Weijia Shi, Yike Wang, Wenxuan Ding, Vidhisha Balachandran, and Yulia Tsvetkov. Don’t hallucinate, abstain: Identifying LLM knowledge gaps via multi-LLM collaboration. In Proceedings of ACL, 2024.arxiv.org/abs/2402.00367
- [9]Yonatan Geifman and Ran El-Yaniv. Selective classification for deep neural networks. In Advances in Neural Information Processing Systems, 2017.arxiv.org/abs/1705.08500
- [10]Robert Geirhos, Jörn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A. Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2020.DOI
- [11]Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and Llama Team, AI @ Meta. The Llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.arxiv.org/abs/2407.21783
- [12]Lynn Hasher, David Goldstein, and Thomas Toppino. Frequency and the conference of referential validity. Journal of Verbal Learning and Verbal Behavior, 16(1):107–112, 1977.DOI
- [13]Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. Constructing a multi-hop QA dataset for comprehensive evaluation of reasoning steps. In Proceedings of COLING, 2020.
- [14]Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022.arxiv.org/abs/2106.09685
- [15]Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan Arik, Dong Wang, Hamed Zamani, and Jiawei Han. Search-R1: Training LLMs to reason and leverage search engines with reinforcement learning. arXiv preprint arXiv:2503.09516, 2025.arxiv.org/abs/2503.09516
- [16]Mandar Joshi, Eunsol Choi, Daniel S. Weld, and Luke Zettlemoyer. TriviaQA: A large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of ACL, 2017.arxiv.org/abs/1705.03551
- [17]Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, Scott Johnston, Sheer El-Showk, Andy Jones, Nelson Elhage, Tristan Hume, Anna Chen, Yuntao Bai, Sam Bowman, Stanislav Fort, Deep Ganguli, Danny Hernandez, Josh Jacobson, Jackson Kernion, Shauna Kravec, Liane Lovitt, Kamal Ndousse, Catherine Olsson, Sam Ringer, Dario Amodei, Tom Brown, Jack Clark, Nicholas Joseph, Ben Mann, Sam McCandlish, Chris Olah, and Jared Kaplan. Language models (mostly) know what they know. arXiv preprint arXiv:2207.05221, 2022.arxiv.org/abs/2207.05221
- [18]Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala, and Edwin Zhang. Why language models hallucinate. arXiv preprint arXiv:2509.04664, 2025.arxiv.org/abs/2509.04664
- [19]Amita Kamath, Robin Jia, and Percy Liang. Selective question answering under domain shift. In Proceedings of ACL, 2020.arxiv.org/abs/2006.09462
- [20]Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. Dense passage retrieval for open-domain question answering. In Proceedings of EMNLP, 2020.arxiv.org/abs/2004.04906
- [21]Divyansh Kaushik, Eduard Hovy, and Zachary Lipton. Learning the difference that makes a difference with counterfactually-augmented data. In International Conference on Learning Representations, 2020.
- [22]Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Ilia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. Natural questions: A benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:452–466, 2019.DOI
- [23]Shayne Longpre, Kartik Perisetla, Anthony Chen, Nikhil Ramesh, Chris DuBois, and Sameer Singh. Entity-based knowledge conflicts in question answering. In Proceedings of EMNLP, 2021.
- [24]Alex Mallen, Akari Asai, Victor Zhong, Rajarshi Das, Daniel Khashabi, and Hannaneh Hajishirzi. When not to trust language models: Investigating effectiveness of parametric and non-parametric memories. In Proceedings of ACL, 2023.
- [25]Jun Nie, Zhiqin Yang, Zhenheng Tang, Yonggang Zhang, Xiaowen Chu, Xinmei Tian, and Bo Han. DRNOISE: Benchmarking deep research agents in misleading evidence environments. arXiv preprint arXiv:2607.17291, 2026.arxiv.org/abs/2607.17291
- [26]Yikang Pan, Liangming Pan, Wenhu Chen, Preslav Nakov, Min-Yen Kan, and William Yang Wang. On the risk of misinformation pollution with large language models. In Findings of EMNLP, 2023.arxiv.org/abs/2305.13661
- [27]Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A. Smith, and Mike Lewis. Measuring and narrowing the compositionality gap in language models. In Findings of EMNLP, 2023.arxiv.org/abs/2210.03350
- [28]Stephen Robertson and Hugo Zaragoza. The probabilistic relevance framework: BM25 and beyond. Foundations and Trends in Information Retrieval, 3(4):333–389, 2009.
- [29]Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. DeepSeekMath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024.arxiv.org/abs/2402.03300
- [30]Yaorui Shi, Sihang Li, Chang Wu, Zhiyuan Liu, Junfeng Fang, Hengxing Cai, An Zhang, and Xiang Wang. Search and refine during think: Facilitating knowledge refinement for improved retrieval-augmented reasoning. In Advances in Neural Information Processing Systems, 2025.arxiv.org/abs/2505.11277
- [31]Huatong Song, Jinhao Jiang, Yingqian Min, Jie Chen, Zhipeng Chen, Wayne Xin Zhao, Lei Fang, and Ji-Rong Wen. R1-Searcher: Incentivizing the search capability in LLMs via reinforcement learning. arXiv preprint arXiv:2503.05592, 2025.arxiv.org/abs/2503.05592
- [32]Linxin Song, Taiwei Shi, and Jieyu Zhao. The hallucination tax of reinforcement finetuning. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 2105–2120, 2025.arxiv.org/abs/2505.13988
- [33]Hao Sun, Zile Qiao, Jiayan Guo, Xuanbo Fan, Yingyan Hou, Yong Jiang, Pengjun Xie, Yan Zhang, Fei Huang, and Jingren Zhou. ZeroSearch: Incentivize the search capability of LLMs without searching. arXiv preprint arXiv:2505.04588, 2025.arxiv.org/abs/2505.04588
- [34]Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. MuSiQue: Multihop questions via single-hop question composition. Transactions of the Association for Computational Linguistics, 2022.arxiv.org/abs/2108.00573
- [35]Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions. In Proceedings of ACL, 2023.arxiv.org/abs/2212.10509
- [36]Han Wang, Archiki Prasad, Elias Stengel-Eskin, and Mohit Bansal. Retrieval-augmented generation with conflicting evidence. arXiv preprint arXiv:2504.13079, 2025.arxiv.org/abs/2504.13079
- [37]Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. Text embeddings by weakly-supervised contrastive pre-training. arXiv preprint arXiv:2212.03533, 2022.arxiv.org/abs/2212.03533
- [38]Sikun Wang, Yixi Zhou, Lei Fan, and Fan Zhang. GraphEcho: Structural redundancy and evidence provenance in LLM graph agents. arXiv preprint arXiv:2609.17695, 2026.arxiv.org/abs/2609.17695
- [39]Bingbing Wen, Jihan Yao, Shangbin Feng, Chenjun Xu, Yulia Tsvetkov, Bill Howe, and Lucy Lu Wang. Know your limits: A survey of abstention in large language models. Transactions of the Association for Computational Linguistics, 13, 2025.arxiv.org/abs/2407.18418
- [40]Kevin Wu, Eric Wu, and James Zou. ClashEval: Quantifying the tug-of-war between an LLM’s internal prior and external evidence. In NeurIPS Datasets and Benchmarks Track, 2024.arxiv.org/abs/2404.10198
- [41]Chong Xiang, Tong Wu, Zexuan Zhong, David Wagner, Danqi Chen, and Prateek Mittal. Certifiably robust RAG against retrieval corruption. arXiv preprint arXiv:2405.15556, 2024.arxiv.org/abs/2405.15556
- [42]Shitao Xiao, Zheng Liu, Peitian Zhang, Niklas Muennighoff, Defu Lian, and Jian-Yun Nie. C-Pack: Packed resources for general chinese embeddings. In Proceedings of SIGIR, 2024.arxiv.org/abs/2309.07597
- [43]Jian Xie, Kai Zhang, Jiangjie Chen, Renze Lou, and Yu Su. Adaptive chameleon or stubborn sloth: Revealing the behavior of large language models in knowledge conflicts. In International Conference on Learning Representations, 2024.arxiv.org/abs/2305.13300
- [44]Rongwu Xu, Zehan Qi, Zhijiang Guo, Cunxiang Wang, Hongru Wang, Yue Zhang, and Wei Xu. Knowledge conflicts for LLMs: A survey. In Proceedings of EMNLP, 2024.
- [45]Shi-Qi Yan, Jia-Chen Gu, Yun Zhu, and Zhen-Hua Ling. Corrective retrieval augmented generation. arXiv preprint arXiv:2401.15884, 2024.arxiv.org/abs/2401.15884
- [46]An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Cheng en Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jianglong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jing Zhou, Jingren Zhou, Junyang Lin, Kai Dang, Keqin Bao, Kexin Yang, Le Yu, Lianghao Deng, Mei Li, Mingfeng Xue, Mingze Li, Pei Zhang, Peng Wang, Qin Zhu, Rui Men, Ruize Gao, Shixuan Liu, Shuang Luo, Tianhao Li, T ianyi Tang, Wenbiao Yin, Xingzhang Ren, Xinyu Wang, Xinyu Zhang, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yinger Zhang, Yu Wan, Yuqiong Liu, Zekun Wang, Zeyu Cui, Zhenru Zhang, Zhipeng Zhou, and Zihan Qiu. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025.arxiv.org/abs/2505.09388
- [47]An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Tianyi Tang, Tingyu Xia, Xingzhang Ren, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yu Wan, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zihan Qiu. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2025.arxiv.org/abs/2412.15115
- [48]Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. HotpotQA: A dataset for diverse, explainable multi-hop question answering. In Proceedings of EMNLP, 2018.arxiv.org/abs/1809.09600
- [49]Fengji Zhang, Tianyu Fan, Yuxiang Zheng, Xinyao Niu, Chengen Huang, Jacky Keung, and Bei Chen. To answer or to abstain: Mitigating search-agent hallucinations via abstention-aware reinforcement learning. arXiv preprint arXiv:2607.10738, 2026.arxiv.org/abs/2607.10738
- [50]Hanning Zhang, Shizhe Diao, Yong Lin, Yi R. Fung, Qing Lian, Xingyao Wang, Yangyi Chen, Heng Ji, and Tong Zhang. R-Tuning: Instructing large language models to say “I don’t know”. In Proceedings of NAACL, 2024.
- [51]Yuxiang Zheng, Dayuan Fu, Xiangkun Hu, Xiaojie Cai, Lyumanshan Ye, Pengrui Lu, and Pengfei Liu. DeepResearch: Scaling deep research via reinforcement learning in real-world environments. In Proceedings of EMNLP, 2025.arxiv.org/abs/2504.03160
- [52]Huichi Zhou, Kin-Hei Lee, Zhonghao Zhan, Yue Chen, Zhenhao Li, Zhaoyang Wang, Hamed Haddadi, and Emine Yilmaz. TrustRAG: Enhancing robustness and trustworthiness in retrieval-augmented generation. arXiv preprint arXiv:2501.00879, 2025.arxiv.org/abs/2501.00879
- [53]Ruihua Zhou, Hao Huang, Yijie Zheng, Jiatang Luo, Yunpeng Li, Yuling Liu, and Dan Du. Evidence laundering: Deep research agents mistake correlated sources for independent evidence. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 11682–11693, 2026.DOI
- [54]Wei Zou, Runpeng Geng, Binghui Wang, and Jinyuan Jia. PoisonedRAG: Knowledge corruption attacks to retrieval-augmented generation of large language models. In USENIX Security Symposium, 2025.arxiv.org/abs/2402.07867




